Industry View · Data Centers & REITs
AI has turned the data center from a real-estate footnote into the single most capital-intensive build-out of the decade. Hyperscaler capex surged sharply in 2025 — with data center capex growing roughly 59% year-on-year in the third quarter alone — and worldwide data center capex is forecast to clear $1 trillion later this decade. Yet vacancy sits near 1.4% and the binding constraint has shifted from chips to megawatts.
Demand, pre-commitment, ~$1T capex; build-out and grid constraint, no demonstrated return
Interconnection queues, PPAs and nuclear offtake now gate every project; whoever secures firm power wins the site.
Single-campus requirements of 1+ GW (Stargate, Crusoe Abilene) turn site selection into a multi-billion-dollar real-estate bet.
30kW+ racks force direct-liquid and immersion cooling; thermal and power gear is the supply-chain chokepoint.
~74% of under-construction capacity is pre-leased before a slab is poured, pushing availability years out.
Neoclouds turn contracted power into multi-year GPU backlogs sold to AI labs and hyperscalers.
01 · The thesis
The demand signal is unambiguous. A large volume of data center capacity was under construction globally through late 2025, the bulk of it in the US, and roughly 74% of capacity under construction is pre-committed — leases or owner-occupied — often for delivery in 2027 and beyond. Record-low vacancy means this is not a 2000-era overbuild: tenants are the most creditworthy companies on earth, and power-interconnection delays are themselves throttling supply faster than any discipline a developer could impose. The risk has migrated from demand to the grid and the balance sheet. PJM came up roughly 6,625 MW short in its December 2025 capacity auction and expects its first capacity shortfall by summer 2027; data center load growth is outpacing new generation roughly 2-to-1. The winners are those who control power, land and interconnection queues — and who can finance multi-gigawatt projects with single-tenant offtake. The exposure is concentration: neoclouds with backlogs leaning on one or two hyperscale counterparties, and ratepayers who may absorb stranded-cost risk if the AI capex curve ever bends.
Interconnection queues, PPAs and nuclear offtake now gate every project; whoever secures firm power wins the site.
Single-campus requirements of 1+ GW (Stargate, Crusoe Abilene) turn site selection into a multi-billion-dollar real-estate bet.
30kW+ racks force direct-liquid and immersion cooling; thermal and power gear is the supply-chain chokepoint.
~74% of under-construction capacity is pre-leased before a slab is poured, pushing availability years out.
Neoclouds turn contracted power into multi-year GPU backlogs sold to AI labs and hyperscalers.
02 · The two clocks
The power clock. S&P Global projects data center grid demand rose sharply in 2025 and roughly triples by 2030; Epoch AI estimates US AI data centers will need 20-30 GW by late 2027. Demand is arriving in gigawatt blocks the grid cannot energize on schedule. The build clock. Vacancy at a record 1.4% and ~74% pre-commitment mean capacity is spoken for years ahead, but power-interconnection and equipment lead times now stretch project timelines well past approval, pushing real delivery risk onto tenants who have already booked the space. The capital clock. With worldwide capex forecast to approach $1 trillion later this decade and hyperscalers each raising spend at double-digit-to-triple-digit rates, the question is no longer whether demand exists but whether single-tenant offtake and debt-heavy neocloud balance sheets can survive any pause in the AI spending curve.
03 · Public players & exposure
We plot the listed players on two editorial axes — how exposed each is to AI disruption, against how ready its data, brand and position are to be the answer. The figures in the table are sourced; the placement is our read.
04 · Private flagships
The companies attacking this industry AI-first, with disclosed funding where available:
Converts contracted Nvidia GPUs and power into multi-year compute backlogs sold to OpenAI, Meta and Microsoft; the purest public proxy for AI infrastructure demand — and its concentration risk.
Builds gigawatt campuses tying power generation to compute; anchor of the Stargate Abilene site and now pushing into modular, grab-and-go data centers.
Supplies the power and liquid-cooling backbone every AI hall now requires; a 2.9x Q4 book-to-bill makes it a leading indicator for the whole pipeline.
OpenAI, SoftBank, Oracle and MGX building toward 10 GW of dedicated AI capacity; the project that recast data centers as national-scale infrastructure.
Repositioning a global colo portfolio toward AI training and inference, with power-bank allocations now the scarce resource it leases.
05 · Signals
06 · The exposure read
AI rewards clean, structured advantage and punishes friction. The line runs through who owns the data, the brand and the customer — and who is merely a step the technology can route around.
Sources
AI-era backlogs and capacity signals (2025)
Source: company filings and IR (Vertiv, CoreWeave, Digital Realty) and Crusoe/Blue Owl JV disclosures, 2025-2026. ($ billion)
PJM capacity auction falls ~6,625 MW short
The largest US grid operator signals it expects its first capacity shortfall by summer 2027 as data center load outpaces new generation 2-to-1.
Vacancy hits record-low 1.4%
CBRE reports North American primary-market vacancy at 1.4% with roughly 74% of under-construction capacity pre-committed.
Nuclear offtake pipeline nearly doubles
The IEA tracks conditional data-center-to-SMR agreements rising from 25 GW (end-2024) to 45 GW as operators chase firm power.
Vertiv orders surge 252%
A 2.9x book-to-bill and $15bn backlog confirm the cooling-and-power supply chain is the live chokepoint of the AI build.
Stargate expands across new sites
OpenAI's JV adds capacity toward its 10 GW target, with hundreds of billions committed across sites.
Converts contracted Nvidia GPUs and power into multi-year compute backlogs sold to OpenAI, Meta and Microsoft; the purest public proxy for AI infrastructure demand — and its concentration risk.
Builds gigawatt campuses tying power generation to compute; anchor of the Stargate Abilene site and now pushing into modular, grab-and-go data centers.
Supplies the power and liquid-cooling backbone every AI hall now requires; a 2.9x Q4 book-to-bill makes it a leading indicator for the whole pipeline.
OpenAI, SoftBank, Oracle and MGX building toward 10 GW of dedicated AI capacity; the project that recast data centers as national-scale infrastructure.
Repositioning a global colo portfolio toward AI training and inference, with power-bank allocations now the scarce resource it leases.
| Company | Stance | The sourced fact |
|---|---|---|
| Equinix EQIX | Defensive incumbent | Posted 2025 revenue of $9.22bn, up ~5%, with Americas cabinets reaching ~154,745, up 7.4% YoY (DCD/S&P Global). |
| Digital Realty DLR | Scaled REIT | Signed $1.2bn of bookings in 2025 — a second straight year above $1bn — for a record ~$1.4bn backlog; AI drove over 50% of bookings (Investing.com). |
| CoreWeave CRWV | Leveraged neocloud | FY2025 revenue of $5.13bn (+168%); revenue backlog of $66.8bn at YE2025 rising to $99.4bn by Q1 2026, but on heavy concentration and capex (SEC/CNBC). |
| Vertiv VRT | Picks-and-shovels | Q4 2025 net sales ~$2.88bn; backlog of $15.0bn (+109% YoY) with organic orders up ~252% (Vertiv IR). |
| Oracle ORCL | Stargate builder | Lead infrastructure partner in the $500bn, 10 GW Stargate JV with OpenAI, SoftBank and MGX (OpenAI). |
| Iron Mountain IRM | Diversifying REIT | Tracked among major colo REITs expanding into AI-grade capacity in a tight US colocation market (DCD). |
| American Tower AMT | Edge optionality | Tracked alongside DLR/EQIX in Q4 2025 colo results as it leans on its CoreSite data center unit for AI demand (DCD). |
| Crusoe CRUSOE | AI-factory developer | Raised a $1.375bn Series E at a ~$10bn valuation; anchors the ~$15bn, 1.2 GW Abilene campus JV with Blue Owl (Crusoe/Bloomberg). |
| Vantage VANTAGE | Hyperscale dev | Developing a multi-gigawatt Stargate-linked campus in Texas, a multi-billion-dollar investment (Blackridge Research). |
| Nebius NBIS | Neocloud upstart | Among the Nvidia-backed neoclouds (with CoreWeave and Nscale) enabling Stargate-scale GPU capacity expansion (DCD). |